Transparencycanon: methodology v1.0.0
Last updated 2026-07-06 · The exact live numbers are published at avieros.com/methodology and read from the ranking code itself.
This page explains how Avieros decides what to show you: the structure of the ranking system, how Candidate Watch scores are built, and the versioning policy that makes silent changes impossible. For the exact live weights and tier tables, always use the canonical methodology page — its numbers are read from the same code that does the ranking, and an automated test fails our build if page and system ever disagree.
1. How search ranking works
Every web result is scored by rank-kernel-v1, our own ranking kernel — never by the discovery provider’s ordering. The score is a weighted blend of five factors. As published at methodology v1.0.0 (2026-07-05):
| Factor | Weight (v1.0.0) | What it measures |
|---|---|---|
| Relevance | 45% | How well the result’s title and snippet match the query. |
| Source trust | 25% | The trust tier of the result’s domain — by rule, not opinion: .gov is always Tier 1, .edu always Tier 2, other domains come from a reviewed registry checked into the codebase, and anything unlisted scores lowest by default. |
| Freshness | 15% | How recent the result is, when the source provides a date. |
| Local relevance | 10% | How confidently the result matches your area when you searched with local intent. |
| Domain diversity | 5% | A penalty when one website already fills multiple slots. |
On top of the score, a hard rule limits any single domain to at most 2 results in the top 10 — no single website can dominate the page.
2. How Candidate Watch scoring works
Candidate Watch publishes what elected officials say and what they do, built on three structural rules:
- Proof-based: every published claim carries a link to a primary source — an official government record, not a summary of a summary. No source, no render.
- Party-neutral by construction: the same scoring rules, the same data sources, and the same rendering apply to every official regardless of party. There is no per-party logic anywhere in the scoring path, and our test data is deliberately party-balanced so imbalances would surface in testing.
- Gaps stay visible: where a record is missing — for example, roll-call floor votes are not yet ingested — the product says so rather than scoring around the hole.
3. Versioning & changelog policy
The methodology is versioned like software, because it governs software:
- Every change ships with a changelog entry. Any change to ranking weights, trust tiers, Candidate Watch scoring, or data sources must include a new entry in the public changelog in the same code change — an automated test fails the build otherwise.
- History cannot be rewritten silently. The full edit history of the changelog is preserved in version control.
- The page cannot drift from the code. The published numbers are read from the ranking code itself and verified by a build-failing sync test.
Methodology v1.0.0, published 2026-07-05. The live version indicator and full changelog are on avieros.com/methodology.
4. Known limitations
Published limitations at v1.0.0 — stated on the canonical page and repeated here because hiding them would defeat the point:
- Roll-call votes are missing. Yea/nay floor votes are not yet ingested, so no say-vs-do comparison is published yet.
- Web discovery depends on a third-party provider. General web results are discovered externally before we re-rank them. Civic and money data do not pass through that provider.
- Candidate coverage is small. Candidate Watch currently covers a seeded set of officials, not every officeholder in the country.
- AI-generated bill summaries are not live. When they ship, each will carry a persistent “AI-generated” label and a direct link to the official bill text.